Project: Build a Traffic Sign Recognition Classifier¶

The present work designs and develops a deep convolutional neural network architecture to classify german traffic signs

Step 1: Load the data and summarize¶

# Load pickled data
import pickle
import numpy as np
import tensorflow as tf


# Fill this in based on where you saved the training and testing data
training_file = "data/train.p"
validation_file= "data/valid.p"
testing_file = "data/test.p"

with open(training_file, mode='rb') as f:
    train = pickle.load(f)
with open(validation_file, mode='rb') as f:
    valid = pickle.load(f)
with open(testing_file, mode='rb') as f:
    test = pickle.load(f)
    
X_train, y_train = train['features'], train['labels']
X_valid, y_valid = valid['features'], valid['labels']
X_test, y_test = test['features'], test['labels']

Dataset summary¶

The pickled data is a dictionary with 4 key/value pairs:

'features' is a 4D array containing raw pixel data of the traffic sign images, (num examples, width, height, channels).
'labels' is a 1D array containing the label/class id of the traffic sign. The file signnames.csv contains id -> name mappings for each id.
'sizes' is a list containing tuples, (width, height) representing the original width and height the image.
'coords' is a list containing tuples, (x1, y1, x2, y2) representing coordinates of a bounding box around the sign in the image. THESE COORDINATES ASSUME THE ORIGINAL IMAGE. THE PICKLED DATA CONTAINS RESIZED VERSIONS (32 by 32) OF THESE IMAGES

Complete the basic data summary below. Use python, numpy and/or pandas methods to calculate the data summary rather than hard coding the results. For example, the pandas shape method might be useful for calculating some of the summary results.

Provide a Basic Summary of the Data Set Using Python, Numpy and/or Pandas¶

# Number of training examples
n_train = len(X_train)

# Number of validation examples
n_validation = len(X_valid)

# Number of testing examples.
n_test = len(X_test)

# What's the shape of an traffic sign image?
image_shape = X_train.shape[1:]

# How many unique classes/labels there are in the dataset.
n_classes = len(set(list(y_test)+list(y_train)+list(y_valid)))

class_labels = []
with open('signnames.txt','r') as f:
    s = f.read() 
    class_labels = s.split('\n')
    class_labels.pop()

print("Number of training examples =", n_train)
print("Number of validation examples =", n_validation)
print("Number of testing examples =", n_test)
print("Image data shape =", image_shape)
print("Number of classes =", n_classes)

Number of training examples = 34799
Number of validation examples = 4410
Number of testing examples = 12630
Image data shape = (32, 32, 3)
Number of classes = 43

Include an exploratory visualization of the dataset¶

Visualize the German Traffic Signs Dataset using the pickled file(s). This is open ended, suggestions include: plotting traffic sign images, plotting the count of each sign, etc.

The Matplotlib examples and gallery pages are a great resource for doing visualizations in Python.

NOTE: It's recommended you start with something simple first. If you wish to do more, come back to it after you've completed the rest of the sections. It can be interesting to look at the distribution of classes in the training, validation and test set. Is the distribution the same? Are there more examples of some classes than others?

import matplotlib.pyplot as plt
# Visualizations will be shown in the notebook.
%matplotlib inline

def random_index(i):
    return np.random.randint(0, i)

rows = 2
cols = 8
fig, ax = plt.subplots(rows,cols, sharex ='col', sharey='row', figsize=(16,5))

for i in range(rows):
    for j in range(cols):
        rand_indx = random_index(n_train-1)
        ax[i,j].imshow(X_train[rand_indx])
        ax[i,j].set_title('Label {}'.format(y_train[rand_indx]))
        ax[i,j].axis('off')  # clear x- and y-axes
fig.suptitle('Overview of training set images', fontsize=16)
plt.show()

Histogram of class image distribution in training set¶

# Draw a histogram of how many features we have per label
# Store the number of images present in the training set for a given label
bincounts = []
# Store the indexes of all the images in the training set for a given label
classes_indexes = []
# Image label vector
x = np.arange(n_classes)

for i in range(n_classes):
    classes_indexes.append(np.where(y_train==i)[0])
    bincounts.append(len(classes_indexes[-1]))
fig, ax = plt.subplots()
fig.suptitle('Image class distribution in training set', fontsize=16)
ax.bar(x, bincounts)
plt.xticks(x[0::2])
plt.show()

#This array stores those classes with a number of samples > 1250
bigc = np.argwhere((np.array(bincounts)>1250)==True)

Representation of images per class label¶

rows = 2
cols = 8

for c in range(len(classes_indexes)):
    fig, ax = plt.subplots(rows,cols, sharex ='col', sharey='row', figsize=(16,5))
    for i in range(rows):
        for j in range(cols):
            rand_indx = np.random.randint(0, len(classes_indexes[c])-1)
            ax[i,j].imshow(X_train[classes_indexes[c][rand_indx]])
            ax[i,j].set_title('Label {}'.format(y_train[classes_indexes[c][rand_indx]]))
            ax[i,j].axis('off')  # clear x- and y-axes
    fig.suptitle('Set of images for class {}: {}'.format(c, class_labels[c]), fontsize=16)
    #plt.savefig('class_{}.png'.format(c))
    plt.show()

Data augmentation¶

The following data augmentation techniques are used in the process:

Cropping
Rotation
Scaling
Translation
~~Gaussian blur~~
~~Contrast normalization~~
~~Additive Gaussian Noise~~

import imgaug as ia
from imgaug import augmenters as iaa
import numpy as np

ia.seed(1)

# Example batch of images.
# The array has shape (?, 32, 32, 3) and dtype uint8.

seq = iaa.Sequential([
    #iaa.Fliplr(0.5), # horizontal flips
    iaa.Crop(percent=(0, 0.1)), # random crops
    # Small gaussian blur with random sigma between 0 and 0.5.
    # But we only blur about 50% of all images.
    #iaa.Sometimes(0.5,
    #    iaa.GaussianBlur(sigma=(0, 0.5))
    #),
    # Strengthen or weaken the contrast in each image.
    #iaa.ContrastNormalization((0.75, 1.5)),
    # Add gaussian noise.
    # For 50% of all images, we sample the noise once per pixel.
    # For the other 50% of all images, we sample the noise per pixel AND
    # channel. This can change the color (not only brightness) of the
    # pixels.
    #iaa.AdditiveGaussianNoise(loc=0, scale=(0.0, 0.05*255), per_channel=0.5),
    # Make some images brighter and some darker.
    # In 20% of all cases, we sample the multiplier once per channel,
    # which can end up changing the color of the images.
    #iaa.Multiply((0.8, 1.2), per_channel=0.2),
    # Apply affine transformations to each image.
    # Scale/zoom them, translate/move them, rotate them and shear them.
    iaa.Affine(
        scale={"x": (0.8, 1.2), "y": (0.8, 1.2)},
        translate_percent={"x": (-0.2, 0.2), "y": (-0.2, 0.2)},
        #shear=(-8, 8),
        rotate=(-15, 15)
    )
], random_order=True) # apply augmenters in random order

print('Image and label set shapes before data set augmentation: {} , {}'.format(X_train.shape, y_train.shape))

Image and label set shapes before data set augmentation: (34799, 32, 32, 3) , (34799,)

from IPython.display import clear_output, display
print('Starting image augumentation...\n')
for i in range(n_classes):
    if bincounts[i] < 1250:
        cycles = 2000//bincounts[i]
        for counter in range(cycles):
            class_img = X_train[classes_indexes[i]]
            y_img = y_train[classes_indexes[i]]
            augmented_images = seq.augment_images(class_img)
            X_train = np.concatenate((X_train, augmented_images))
            y_train = np.concatenate((y_train, y_img))
    else:
        class_img = X_train[classes_indexes[i]][0:1000]
        y_img = y_train[classes_indexes[i]][0:1000]
        augmented_images = seq.augment_images(class_img)
        X_train = np.concatenate((X_train, augmented_images))
        y_train = np.concatenate((y_train, y_img))
    clear_output(wait=True)
    print('Augmentation of class {} completed \n'.format(i))

print('Image and label set shapes after data set augmentation: {} , {}'.format(X_train.shape, y_train.shape))

Augmentation of class 42 completed 

Image and label set shapes after data set augmentation: (100656, 32, 32, 3) , (100656,)

# Draw a histogram of how many features we have per label
# Store the number of images present in the training set for a given label
bincounts = []
# Store the indexes of all the images in the training set for a given label
classes_indexes = []
# Image label vector|
x = np.arange(n_classes)

for i in range(n_classes):
    classes_indexes.append(np.where(y_train==i)[0])
    bincounts.append(len(classes_indexes[-1]))
fig, ax = plt.subplots()
fig.suptitle('Image class distribution in training set', fontsize=16)
ax.bar(x, bincounts)
plt.xticks(x[0::2])
plt.show()

Representation of images per class label after augmentation¶

rows = 2
cols = 8

for c in range(len(classes_indexes)):
    fig, ax = plt.subplots(rows,cols, sharex ='col', sharey='row', figsize=(16,5))
    for i in range(rows):
        for j in range(cols):
            rand_indx = np.random.randint(0, len(classes_indexes[c])-1)
            ax[i,j].imshow(X_train[classes_indexes[c][rand_indx]], cmap='gray')
            ax[i,j].set_title('Label {}'.format(y_train[classes_indexes[c][rand_indx]]))
            ax[i,j].axis('off')  # clear x- and y-axes
    fig.suptitle('Set of images for class {}: {}'.format(c, class_labels[c]), fontsize=16)
    plt.show()

Pre-process the Data Set (normalization, grayscale, etc.)¶

Minimally, the image data should be normalized so that the data has mean zero and equal variance. For image data, (pixel - 128)/ 128 is a quick way to approximately normalize the data and can be used in this project.

Other pre-processing steps are optional. You can try different techniques to see if it improves performance.

Use the code cell (or multiple code cells, if necessary) to implement the first step of your project.

from sklearn.utils import shuffle

# Suffle input/outputs so the starting point of data structure is random
X_train, y_train = shuffle(X_train, y_train)
X_valid, y_valid = shuffle(X_valid, y_valid)
X_test, y_test = shuffle(X_test, y_test)

Preprocessing methods¶

Two preprocessing methods were used in this project. The following blocks describe and implement each method.

1. Grayscale and normalization¶

In this method we convert the original image to grayscale using the numpy mean method and normalize the values to be between -1 and 1.

def preprocessing_grayscale(X):
    # Converting RGB images to GRAY scale
    X_gray = X.mean(axis=-1, keepdims=0)
    # Add one dimension to the tensor
    X_gray = X_gray[:,:,:, np.newaxis]
    # Input normalization from values between 0-255 to -1-1
    X_gray = (X_gray-128.0)/128.0
    return X_gray

2. Channel Y of YVU color space, with local/global contrast normalization¶

import numpy
import scipy
import scipy.misc
from PIL import Image
import skimage as sk


def global_contrast_normalization(X, s, lmda, epsilon):

    # replacement for the loop
    X_average = numpy.mean(X)

    X = X - X_average

    # `sum` is here the mean, instead of the sum
    contrast = numpy.sqrt(lmda + numpy.mean(X**2))

    X = s * X / max(contrast, epsilon)

    return X


def preprocessing_channelY(X):
    images =[]
    # convert image to YUV space, only use the Y channel
    #X=0.299*X[:,:,:,0]+0.587*X[:,:,:,1]+0.114*X[:,:,:,2]
    #X=(X/255.0).astype(np.float32)
    
    for i in range(len(X)):
        y = sk.color.rgb2yuv(X[i])[:,:,0]
        if (i%1000==0):
            print('{}th iteration\n'.format(i))
        # Adaptive Equalization
        y = sk.exposure.equalize_adapthist(y, clip_limit=0.03)
        y = global_contrast_normalization(y, 1, 10, 0.000000001)
        images.append(y[:,:,np.newaxis])
    return np.array(images)

The preprocessing available methods are: grayscale_normalization and channelY_normalization. The one to be used is defined in the preprocessing_method flag.

# Dictionary storing both preprocessing functions
preprocessing = {'grayscale_normalization': preprocessing_grayscale ,'channelY_normalization': preprocessing_channelY}

# Choose one between 'grayscale_normalization' and 'channelY_normalization' 
preprocessing_method = 'channelY_normalization'

X_train_pp = preprocessing[preprocessing_method](X_train)
X_valid_pp = preprocessing[preprocessing_method](X_valid)
X_test_pp =  preprocessing[preprocessing_method](X_test)

0th iteration

C:\Users\juand\Anaconda3\envs\tensorflow\lib\site-packages\skimage\util\dtype.py:141: UserWarning: Possible precision loss when converting from float64 to uint16
  .format(dtypeobj_in, dtypeobj_out))

1000th iteration

2000th iteration

3000th iteration

4000th iteration

5000th iteration

6000th iteration

7000th iteration

8000th iteration

9000th iteration

10000th iteration

11000th iteration

12000th iteration

13000th iteration

14000th iteration

15000th iteration

16000th iteration

17000th iteration

18000th iteration

19000th iteration

20000th iteration

21000th iteration

22000th iteration

23000th iteration

24000th iteration

25000th iteration

26000th iteration

27000th iteration

28000th iteration

29000th iteration

30000th iteration

31000th iteration

32000th iteration

33000th iteration

34000th iteration

35000th iteration

36000th iteration

37000th iteration

38000th iteration

39000th iteration

40000th iteration

41000th iteration

42000th iteration

43000th iteration

44000th iteration

45000th iteration

46000th iteration

47000th iteration

48000th iteration

49000th iteration

50000th iteration

51000th iteration

52000th iteration

53000th iteration

54000th iteration

55000th iteration

56000th iteration

57000th iteration

58000th iteration

59000th iteration

60000th iteration

61000th iteration

62000th iteration

63000th iteration

64000th iteration

65000th iteration

66000th iteration

67000th iteration

68000th iteration

69000th iteration

70000th iteration

71000th iteration

72000th iteration

73000th iteration

74000th iteration

75000th iteration

76000th iteration

77000th iteration

78000th iteration

79000th iteration

80000th iteration

81000th iteration

82000th iteration

83000th iteration

84000th iteration

85000th iteration

86000th iteration

87000th iteration

88000th iteration

89000th iteration

90000th iteration

91000th iteration

92000th iteration

93000th iteration

94000th iteration

95000th iteration

96000th iteration

97000th iteration

98000th iteration

99000th iteration

100000th iteration

0th iteration

1000th iteration

2000th iteration

3000th iteration

4000th iteration

0th iteration

1000th iteration

2000th iteration

3000th iteration

4000th iteration

5000th iteration

6000th iteration

7000th iteration

8000th iteration

9000th iteration

10000th iteration

11000th iteration

12000th iteration

# Uncomment this block if you also want to preprocess with the other method. Grayscale normalization in this case
'''
X_train_gray = preprocessing['grayscale_normalization'](X_train)
X_valid_gray = preprocessing['grayscale_normalization'](X_valid)
X_test_gray =  preprocessing['grayscale_normalization'](X_test)
'''

### Preprocess the data here. It is required to normalize the data. Other preprocessing steps could include 
### converting to grayscale, etc.
### Feel free to use as many code cells as needed.

img_index = random_index(X_train.shape[0])
rows = 1
cols = 3

new_image = sk.color.rgb2yuv(X_train[img_index])
color_channel = ['Y', 'U', 'V']

# Image preprocessing of the gray scale version of the image
gray = X_train[img_index].mean(axis=-1,keepdims=0)
gray_n = (gray-128)/128


# Image preprocessing of the y channel version of the image
# Channel y of the image in img_index
y= new_image[:,:,0]
# Local contrast normalization w/ Adaptive Equalization
y = sk.exposure.equalize_adapthist(y, clip_limit=0.03)
# Global contrast normalization
y = global_contrast_normalization(y, 1, 10, 0.000000001)

# Show the different channel of the RGB image in index img_index
fig0, ax0 = plt.subplots(1,2, sharex ='col', sharey='row', figsize=(16,5))
ax0[0].imshow(X_train[img_index])
ax0[0].set_title('Original image')
ax0[0].axis('off')  # clear x- and y-axes

ax0[1].imshow(gray,cmap='gray')
ax0[1].set_title('Grayscale image')
ax0[1].axis('off')  # clear x- and y-axes
fig0.suptitle('Grayscale image', fontsize=16)
plt.show()


# Show the different channel of the RGB image in index img_index
fig, ax = plt.subplots(rows,cols, sharex ='col', sharey='row', figsize=(16,5))
for i in range(rows):
    for j in range(cols): 
        ax[j].imshow(new_image[:,:,j], cmap='gray')
        ax[j].set_title('Color channel {}'.format(color_channel[j]))
        ax[j].axis('off')  # clear x- and y-axes
fig.suptitle('Color space YUV', fontsize=16)
plt.show()




# Show the original image and its preprocessing after gray scale and y channel conditioning
titles = ['Original image', 'Grayscale with normalization', 'Channnel Y with global/local contrast normalization']
fig1, ax1 = plt.subplots(rows,cols, sharex ='col', sharey='row', figsize=(16,5))

ax1[0].imshow(X_train[img_index])
ax1[0].set_title('{}'.format(titles[0]))
ax1[0].axis('off')  # clear x- and y-axes

ax1[1].imshow(gray,cmap='gray')
ax1[1].set_title('{}'.format(titles[1]))
ax1[1].axis('off')  # clear x- and y-axes

ax1[2].imshow(y,cmap='gray')
ax1[2].set_title('{}'.format(titles[2]))
ax1[2].axis('off')  # clear x- and y-axes

fig1.suptitle('Preprocessing image methods', fontsize=16)
plt.show()

from skimage import data, img_as_float
from skimage import exposure

def plot_img_and_hist(image, axes, bins=256):
    """Plot an image along with its histogram and cumulative histogram.

    """
    image = img_as_float(image)
    ax_img, ax_hist = axes
    ax_cdf = ax_hist.twinx()

    # Display image
    ax_img.imshow(image, cmap=plt.cm.gray)
    ax_img.set_axis_off()

    # Display histogram
    ax_hist.hist(image.ravel(), bins=bins, histtype='step', color='black')
    ax_hist.ticklabel_format(axis='y', style='scientific', scilimits=(0, 0))
    ax_hist.set_xlabel('Pixel intensity')
    ax_hist.set_xlim(0, 1)
    ax_hist.set_yticks([])

    # Display cumulative distribution
    img_cdf, bins = exposure.cumulative_distribution(image, bins)
    ax_cdf.plot(bins, img_cdf, 'r')
    ax_cdf.set_yticks([])

    return ax_img, ax_hist, ax_cdf



# Display results
fig = plt.figure(figsize=(10, 5))
axes = np.zeros((2, 2), dtype=np.object)
axes[0, 0] = fig.add_subplot(2, 2, 1)
for i in range(1, 2):
    axes[0, i] = fig.add_subplot(2, 2, 1+i, sharex=axes[0,0], sharey=axes[0,0])
for i in range(0, 2):
    axes[1, i] = fig.add_subplot(2, 2, 3+i)

ax_img, ax_hist, ax_cdf = plot_img_and_hist(gray, axes[:, 0])
ax_img.set_title('Grayscale with normalization')

y_min, y_max = ax_hist.get_ylim()
ax_hist.set_ylabel('Number of pixels')
ax_hist.set_yticks(np.linspace(0, y_max, 5))

ax_img, ax_hist, ax_cdf = plot_img_and_hist(y, axes[:, 1])
ax_img.set_title('Channel Y with local/global contrast normalization')

ax_cdf.set_ylabel('Fraction of total intensity')
ax_cdf.set_yticks(np.linspace(0, 1, 5))

# prevent overlap of y-axis labels
fig.tight_layout()
plt.show()

Plot some of the new images before and after augmentation¶

rows = 2
cols = 8

fig, ax = plt.subplots(rows,cols, sharex ='col', sharey='row', figsize=(16,5))

for i in range(rows):
    for j in range(cols):
        rand_indx = random_index(X_train_pp.shape[0]-1)
        ax[i,j].imshow(X_train_pp[rand_indx][:,:,0], cmap='gray')
        ax[i,j].set_title('Label {}'.format(y_train[rand_indx]))
        ax[i,j].axis('off')  # clear x- and y-axes
fig.suptitle('Training set augmented images with {} preprocessing'.format(preprocessing_method), fontsize=16)
plt.show()

Save augmented and preprocessed training, validation and testing data¶

In this block of code we store the augmented and preprocessed training, validation and testing data. Two methods of preprocessing were used in this project:

The first one takes a color image, converts it to gray scale taking the mean of its pixel values and then normalizes them to be between -1 and 1.
The second method uses local and global contrast normalization over the Y color channel of the YVU color space of the original image.

training_aug_pp = {'features_y': X_train_pp,'features_gray': X_train_gray, 'labels': y_train}
valid_pp = {'features_y': X_valid_pp, 'features_gray': X_valid_gray, 'labels': y_valid}
test_pp = {'features_y': X_test_pp, 'features_gray': X_test_gray, 'labels': y_test}


training_file_aug_pp = "data/train_aug_pp.p"
valid_file_pp = "data/valid_pp.p"
test_file_pp = "data/test_pp.p"


with open(training_file_aug_pp, mode='wb') as f:
    pickle.dump(training_aug_pp, f, protocol=pickle.HIGHEST_PROTOCOL)

with open(valid_file_pp, mode='wb') as f:
    pickle.dump(valid_pp, f, protocol=pickle.HIGHEST_PROTOCOL)

with open(test_file_pp, mode='wb') as f:
    pickle.dump(test_pp, f, protocol=pickle.HIGHEST_PROTOCOL)

Load preprocessed data¶

Load data when a preprocessed step has been previously performed. It will load the data in dictionaries with keys:

'features_y': for features preprocessed in the y channel color and local/global contrast normalization
'features_gray': for features preprocessed in the grayscale version of the original image and pixel intensity normalization

import pickle 
import tensorflow as tf
from sklearn.utils import shuffle

training_file_aug_pp = "data/train_aug_pp.p"
validation_file_pp = "data/valid_pp.p"
testing_file_pp = "data/test_pp.p"

with open(training_file_aug_pp, mode='rb') as f:
    train_pp = pickle.load(f)
with open(validation_file_pp, mode='rb') as f:
    valid_pp = pickle.load(f)
with open(testing_file_pp, mode='rb') as f:
    test_pp = pickle.load(f)

# Choose between two options: 'features_y' and 'features_gray'
input_method = 'features_y'

X_train_pp, y_train = train_pp[input_method],  train_pp['labels']
X_valid_pp, y_valid = valid_pp[input_method], valid_pp['labels']
X_test_pp, y_test = test_pp[input_method], test_pp['labels']

# How many unique classes/labels there are in the dataset.
n_classes = len(set(list(y_test)+list(y_train)+list(y_valid)))

Step 2: Design and Test a Model Architecture¶

Design and implement a deep learning model that learns to recognize traffic signs. Train and test your model on the German Traffic Sign Dataset.

The LeNet-5 implementation shown in the classroom at the end of the CNN lesson is a solid starting point. You'll have to change the number of classes and possibly the preprocessing, but aside from that it's plug and play!

With the LeNet-5 solution from the lecture, you should expect a validation set accuracy of about 0.89. To meet specifications, the validation set accuracy will need to be at least 0.93. It is possible to get an even higher accuracy, but 0.93 is the minimum for a successful project submission.

There are various aspects to consider when thinking about this problem:

Neural network architecture (is the network over or underfitting?)
Play around preprocessing techniques (normalization, rgb to grayscale, etc)
Number of examples per label (some have more than others).
Generate fake data.

Here is an example of a published baseline model on this problem. It's not required to be familiar with the approach used in the paper but, it's good practice to try to read papers like these.

Model Architecture¶

Input¶

This neural network accepts a 32x32x1 image as input, where the final dimension is the number of color channels. Architecture